Tensorflow2 offers roughly 3 code styles of model implementation: Sequential, Functional and Subclassing. This project adopts each style to implement different image classifier architecture. Specifically, three files in models package implements:

• feed forward architecture using Sequential API
• plain CNN architecture using functional API
• ResNet architecture using subclassing

This application is scripted and tested on M2 Apple Silicon machine.

conda env create
conda activate cifar10

User can try out this demo in two ways.

Each of three different pretrained models and corresponding CSV files of logs generated during training process are saved in my S3 repository. By executing main.py script using remote command, those objects can be downloaded, and you are ready to go.

python main.py remote

Optionally, to train and save models in local machine, execute main.py script using local command. Trained models and history files will be saved in directory specified in LOCAL_DIR variable defined in main.py.

python main.py local

To try out trained models, execute Streamlit demo as following.

streamlit run demo.py